Alibaba unveiled Qwen3.5 on Monday, the latest generation of its flagship AI model family, timed to the eve of the Lunar New Year and capping an extraordinary week of new model releases from China's leading AI developers. Qwen3.5 is designed for the agentic AI era — built to take actions across mobile apps, desktop interfaces, and complex multi-step workflows without requiring constant human input.
The model uses a hybrid Mixture-of-Experts architecture that activates only 17 billion parameters per forward pass out of a total of 397 billion, delivering high capability at substantially lower inference cost. Alibaba claims Qwen3.5 is 60% cheaper to run and eight times faster at processing large workloads than its predecessor, while outperforming OpenAI's GPT-5.2, Anthropic's Claude Opus 4.5, and Google's Gemini 3 on key benchmarks — though these figures are self-reported and have not been independently verified.
The model supports 201 languages and dialects, up from 82 in the previous generation, and can process inputs including up to two hours of video. It arrives as both a hosted version on Alibaba's servers and an open-weight release that developers can download, fine-tune, and run on their own infrastructure. The Qwen family has now exceeded 20 million downloads across its versions.